git.oblomov.eu Git - git/blob - pack-write.c

   1 #include "cache.h"
   2 #include "pack.h"
   3 #include "csum-file.h"
   4
   5 void reset_pack_idx_option(struct pack_idx_option *opts)
   6 {
   7         memset(opts, 0, sizeof(*opts));
   8         opts->version = 2;
   9         opts->off32_limit = 0x7fffffff;
  10 }
  11
  12 static int sha1_compare(const void *_a, const void *_b)
  13 {
  14         struct pack_idx_entry *a = *(struct pack_idx_entry **)_a;
  15         struct pack_idx_entry *b = *(struct pack_idx_entry **)_b;
  16         return oidcmp(&a->oid, &b->oid);
  17 }
  18
  19 static int cmp_uint32(const void *a_, const void *b_)
  20 {
  21         uint32_t a = *((uint32_t *)a_);
  22         uint32_t b = *((uint32_t *)b_);
  23
  24         return (a < b) ? -1 : (a != b);
  25 }
  26
  27 static int need_large_offset(off_t offset, const struct pack_idx_option *opts)
  28 {
  29         uint32_t ofsval;
  30
  31         if ((offset >> 31) || (opts->off32_limit < offset))
  32                 return 1;
  33         if (!opts->anomaly_nr)
  34                 return 0;
  35         ofsval = offset;
  36         return !!bsearch(&ofsval, opts->anomaly, opts->anomaly_nr,
  37                          sizeof(ofsval), cmp_uint32);
  38 }
  39
  40 /*
  41  * The *sha1 contains the pack content SHA1 hash.
  42  * The objects array passed in will be sorted by SHA1 on exit.
  43  */
  44 const char *write_idx_file(const char *index_name, struct pack_idx_entry **objects,
  45                            int nr_objects, const struct pack_idx_option *opts,
  46                            const unsigned char *sha1)
  47 {
  48         struct hashfile *f;
  49         struct pack_idx_entry **sorted_by_sha, **list, **last;
  50         off_t last_obj_offset = 0;
  51         int i, fd;
  52         uint32_t index_version;
  53
  54         if (nr_objects) {
  55                 sorted_by_sha = objects;
  56                 list = sorted_by_sha;
  57                 last = sorted_by_sha + nr_objects;
  58                 for (i = 0; i < nr_objects; ++i) {
  59                         if (objects[i]->offset > last_obj_offset)
  60                                 last_obj_offset = objects[i]->offset;
  61                 }
  62                 QSORT(sorted_by_sha, nr_objects, sha1_compare);
  63         }
  64         else
  65                 sorted_by_sha = list = last = NULL;
  66
  67         if (opts->flags & WRITE_IDX_VERIFY) {
  68                 assert(index_name);
  69                 f = hashfd_check(index_name);
  70         } else {
  71                 if (!index_name) {
  72                         struct strbuf tmp_file = STRBUF_INIT;
  73                         fd = odb_mkstemp(&tmp_file, "pack/tmp_idx_XXXXXX");
  74                         index_name = strbuf_detach(&tmp_file, NULL);
  75                 } else {
  76                         unlink(index_name);
  77                         fd = open(index_name, O_CREAT|O_EXCL|O_WRONLY, 0600);
  78                         if (fd < 0)
  79                                 die_errno("unable to create '%s'", index_name);
  80                 }
  81                 f = hashfd(fd, index_name);
  82         }
  83
  84         /* if last object's offset is >= 2^31 we should use index V2 */
  85         index_version = need_large_offset(last_obj_offset, opts) ? 2 : opts->version;
  86
  87         /* index versions 2 and above need a header */
  88         if (index_version >= 2) {
  89                 struct pack_idx_header hdr;
  90                 hdr.idx_signature = htonl(PACK_IDX_SIGNATURE);
  91                 hdr.idx_version = htonl(index_version);
  92                 hashwrite(f, &hdr, sizeof(hdr));
  93         }
  94
  95         /*
  96          * Write the first-level table (the list is sorted,
  97          * but we use a 256-entry lookup to be able to avoid
  98          * having to do eight extra binary search iterations).
  99          */
 100         for (i = 0; i < 256; i++) {
 101                 struct pack_idx_entry **next = list;
 102                 while (next < last) {
 103                         struct pack_idx_entry *obj = *next;
 104                         if (obj->oid.hash[0] != i)
 105                                 break;
 106                         next++;
 107                 }
 108                 hashwrite_be32(f, next - sorted_by_sha);
 109                 list = next;
 110         }
 111
 112         /*
 113          * Write the actual SHA1 entries..
 114          */
 115         list = sorted_by_sha;
 116         for (i = 0; i < nr_objects; i++) {
 117                 struct pack_idx_entry *obj = *list++;
 118                 if (index_version < 2)
 119                         hashwrite_be32(f, obj->offset);
 120                 hashwrite(f, obj->oid.hash, the_hash_algo->rawsz);
 121                 if ((opts->flags & WRITE_IDX_STRICT) &&
 122                     (i && oideq(&list[-2]->oid, &obj->oid)))
 123                         die("The same object %s appears twice in the pack",
 124                             oid_to_hex(&obj->oid));
 125         }
 126
 127         if (index_version >= 2) {
 128                 unsigned int nr_large_offset = 0;
 129
 130                 /* write the crc32 table */
 131                 list = sorted_by_sha;
 132                 for (i = 0; i < nr_objects; i++) {
 133                         struct pack_idx_entry *obj = *list++;
 134                         hashwrite_be32(f, obj->crc32);
 135                 }
 136
 137                 /* write the 32-bit offset table */
 138                 list = sorted_by_sha;
 139                 for (i = 0; i < nr_objects; i++) {
 140                         struct pack_idx_entry *obj = *list++;
 141                         uint32_t offset;
 142
 143                         offset = (need_large_offset(obj->offset, opts)
 144                                   ? (0x80000000 | nr_large_offset++)
 145                                   : obj->offset);
 146                         hashwrite_be32(f, offset);
 147                 }
 148
 149                 /* write the large offset table */
 150                 list = sorted_by_sha;
 151                 while (nr_large_offset) {
 152                         struct pack_idx_entry *obj = *list++;
 153                         uint64_t offset = obj->offset;
 154                         uint32_t split[2];
 155
 156                         if (!need_large_offset(offset, opts))
 157                                 continue;
 158                         split[0] = htonl(offset >> 32);
 159                         split[1] = htonl(offset & 0xffffffff);
 160                         hashwrite(f, split, 8);
 161                         nr_large_offset--;
 162                 }
 163         }
 164
 165         hashwrite(f, sha1, the_hash_algo->rawsz);
 166         finalize_hashfile(f, NULL, CSUM_HASH_IN_STREAM | CSUM_CLOSE |
 167                                     ((opts->flags & WRITE_IDX_VERIFY)
 168                                     ? 0 : CSUM_FSYNC));
 169         return index_name;
 170 }
 171
 172 off_t write_pack_header(struct hashfile *f, uint32_t nr_entries)
 173 {
 174         struct pack_header hdr;
 175
 176         hdr.hdr_signature = htonl(PACK_SIGNATURE);
 177         hdr.hdr_version = htonl(PACK_VERSION);
 178         hdr.hdr_entries = htonl(nr_entries);
 179         hashwrite(f, &hdr, sizeof(hdr));
 180         return sizeof(hdr);
 181 }
 182
 183 /*
 184  * Update pack header with object_count and compute new SHA1 for pack data
 185  * associated to pack_fd, and write that SHA1 at the end.  That new SHA1
 186  * is also returned in new_pack_sha1.
 187  *
 188  * If partial_pack_sha1 is non null, then the SHA1 of the existing pack
 189  * (without the header update) is computed and validated against the
 190  * one provided in partial_pack_sha1.  The validation is performed at
 191  * partial_pack_offset bytes in the pack file.  The SHA1 of the remaining
 192  * data (i.e. from partial_pack_offset to the end) is then computed and
 193  * returned in partial_pack_sha1.
 194  *
 195  * Note that new_pack_sha1 is updated last, so both new_pack_sha1 and
 196  * partial_pack_sha1 can refer to the same buffer if the caller is not
 197  * interested in the resulting SHA1 of pack data above partial_pack_offset.
 198  */
 199 void fixup_pack_header_footer(int pack_fd,
 200                          unsigned char *new_pack_hash,
 201                          const char *pack_name,
 202                          uint32_t object_count,
 203                          unsigned char *partial_pack_hash,
 204                          off_t partial_pack_offset)
 205 {
 206         int aligned_sz, buf_sz = 8 * 1024;
 207         git_hash_ctx old_hash_ctx, new_hash_ctx;
 208         struct pack_header hdr;
 209         char *buf;
 210         ssize_t read_result;
 211
 212         the_hash_algo->init_fn(&old_hash_ctx);
 213         the_hash_algo->init_fn(&new_hash_ctx);
 214
 215         if (lseek(pack_fd, 0, SEEK_SET) != 0)
 216                 die_errno("Failed seeking to start of '%s'", pack_name);
 217         read_result = read_in_full(pack_fd, &hdr, sizeof(hdr));
 218         if (read_result < 0)
 219                 die_errno("Unable to reread header of '%s'", pack_name);
 220         else if (read_result != sizeof(hdr))
 221                 die_errno("Unexpected short read for header of '%s'",
 222                           pack_name);
 223         if (lseek(pack_fd, 0, SEEK_SET) != 0)
 224                 die_errno("Failed seeking to start of '%s'", pack_name);
 225         the_hash_algo->update_fn(&old_hash_ctx, &hdr, sizeof(hdr));
 226         hdr.hdr_entries = htonl(object_count);
 227         the_hash_algo->update_fn(&new_hash_ctx, &hdr, sizeof(hdr));
 228         write_or_die(pack_fd, &hdr, sizeof(hdr));
 229         partial_pack_offset -= sizeof(hdr);
 230
 231         buf = xmalloc(buf_sz);
 232         aligned_sz = buf_sz - sizeof(hdr);
 233         for (;;) {
 234                 ssize_t m, n;
 235                 m = (partial_pack_hash && partial_pack_offset < aligned_sz) ?
 236                         partial_pack_offset : aligned_sz;
 237                 n = xread(pack_fd, buf, m);
 238                 if (!n)
 239                         break;
 240                 if (n < 0)
 241                         die_errno("Failed to checksum '%s'", pack_name);
 242                 the_hash_algo->update_fn(&new_hash_ctx, buf, n);
 243
 244                 aligned_sz -= n;
 245                 if (!aligned_sz)
 246                         aligned_sz = buf_sz;
 247
 248                 if (!partial_pack_hash)
 249                         continue;
 250
 251                 the_hash_algo->update_fn(&old_hash_ctx, buf, n);
 252                 partial_pack_offset -= n;
 253                 if (partial_pack_offset == 0) {
 254                         unsigned char hash[GIT_MAX_RAWSZ];
 255                         the_hash_algo->final_fn(hash, &old_hash_ctx);
 256                         if (!hasheq(hash, partial_pack_hash))
 257                                 die("Unexpected checksum for %s "
 258                                     "(disk corruption?)", pack_name);
 259                         /*
 260                          * Now let's compute the SHA1 of the remainder of the
 261                          * pack, which also means making partial_pack_offset
 262                          * big enough not to matter anymore.
 263                          */
 264                         the_hash_algo->init_fn(&old_hash_ctx);
 265                         partial_pack_offset = ~partial_pack_offset;
 266                         partial_pack_offset -= MSB(partial_pack_offset, 1);
 267                 }
 268         }
 269         free(buf);
 270
 271         if (partial_pack_hash)
 272                 the_hash_algo->final_fn(partial_pack_hash, &old_hash_ctx);
 273         the_hash_algo->final_fn(new_pack_hash, &new_hash_ctx);
 274         write_or_die(pack_fd, new_pack_hash, the_hash_algo->rawsz);
 275         fsync_or_die(pack_fd, pack_name);
 276 }
 277
 278 char *index_pack_lockfile(int ip_out)
 279 {
 280         char packname[GIT_MAX_HEXSZ + 6];
 281         const int len = the_hash_algo->hexsz + 6;
 282
 283         /*
 284          * The first thing we expect from index-pack's output
 285          * is "pack\t%40s\n" or "keep\t%40s\n" (46 bytes) where
 286          * %40s is the newly created pack SHA1 name.  In the "keep"
 287          * case, we need it to remove the corresponding .keep file
 288          * later on.  If we don't get that then tough luck with it.
 289          */
 290         if (read_in_full(ip_out, packname, len) == len && packname[len-1] == '\n') {
 291                 const char *name;
 292                 packname[len-1] = 0;
 293                 if (skip_prefix(packname, "keep\t", &name))
 294                         return xstrfmt("%s/pack/pack-%s.keep",
 295                                        get_object_directory(), name);
 296         }
 297         return NULL;
 298 }
 299
 300 /*
 301  * The per-object header is a pretty dense thing, which is
 302  *  - first byte: low four bits are "size", then three bits of "type",
 303  *    and the high bit is "size continues".
 304  *  - each byte afterwards: low seven bits are size continuation,
 305  *    with the high bit being "size continues"
 306  */
 307 int encode_in_pack_object_header(unsigned char *hdr, int hdr_len,
 308                                  enum object_type type, uintmax_t size)
 309 {
 310         int n = 1;
 311         unsigned char c;
 312
 313         if (type < OBJ_COMMIT || type > OBJ_REF_DELTA)
 314                 die("bad type %d", type);
 315
 316         c = (type << 4) | (size & 15);
 317         size >>= 4;
 318         while (size) {
 319                 if (n == hdr_len)
 320                         die("object size is too enormous to format");
 321                 *hdr++ = c | 0x80;
 322                 c = size & 0x7f;
 323                 size >>= 7;
 324                 n++;
 325         }
 326         *hdr = c;
 327         return n;
 328 }
 329
 330 struct hashfile *create_tmp_packfile(char **pack_tmp_name)
 331 {
 332         struct strbuf tmpname = STRBUF_INIT;
 333         int fd;
 334
 335         fd = odb_mkstemp(&tmpname, "pack/tmp_pack_XXXXXX");
 336         *pack_tmp_name = strbuf_detach(&tmpname, NULL);
 337         return hashfd(fd, *pack_tmp_name);
 338 }
 339
 340 void finish_tmp_packfile(struct strbuf *name_buffer,
 341                          const char *pack_tmp_name,
 342                          struct pack_idx_entry **written_list,
 343                          uint32_t nr_written,
 344                          struct pack_idx_option *pack_idx_opts,
 345                          unsigned char hash[])
 346 {
 347         const char *idx_tmp_name;
 348         int basename_len = name_buffer->len;
 349
 350         if (adjust_shared_perm(pack_tmp_name))
 351                 die_errno("unable to make temporary pack file readable");
 352
 353         idx_tmp_name = write_idx_file(NULL, written_list, nr_written,
 354                                       pack_idx_opts, hash);
 355         if (adjust_shared_perm(idx_tmp_name))
 356                 die_errno("unable to make temporary index file readable");
 357
 358         strbuf_addf(name_buffer, "%s.pack", hash_to_hex(hash));
 359
 360         if (rename(pack_tmp_name, name_buffer->buf))
 361                 die_errno("unable to rename temporary pack file");
 362
 363         strbuf_setlen(name_buffer, basename_len);
 364
 365         strbuf_addf(name_buffer, "%s.idx", hash_to_hex(hash));
 366         if (rename(idx_tmp_name, name_buffer->buf))
 367                 die_errno("unable to rename temporary index file");
 368
 369         strbuf_setlen(name_buffer, basename_len);
 370
 371         free((void *)idx_tmp_name);
 372 }